xref: /openbmc/linux/block/blk-mq-tag.c (revision 206a81c1)
1 /*
2  * Fast and scalable bitmap tagging variant. Uses sparser bitmaps spread
3  * over multiple cachelines to avoid ping-pong between multiple submitters
4  * or submitter and completer. Uses rolling wakeups to avoid falling of
5  * the scaling cliff when we run out of tags and have to start putting
6  * submitters to sleep.
7  *
8  * Uses active queue tracking to support fairer distribution of tags
9  * between multiple submitters when a shared tag map is used.
10  *
11  * Copyright (C) 2013-2014 Jens Axboe
12  */
13 #include <linux/kernel.h>
14 #include <linux/module.h>
15 #include <linux/random.h>
16 
17 #include <linux/blk-mq.h>
18 #include "blk.h"
19 #include "blk-mq.h"
20 #include "blk-mq-tag.h"
21 
22 static bool bt_has_free_tags(struct blk_mq_bitmap_tags *bt)
23 {
24 	int i;
25 
26 	for (i = 0; i < bt->map_nr; i++) {
27 		struct blk_align_bitmap *bm = &bt->map[i];
28 		int ret;
29 
30 		ret = find_first_zero_bit(&bm->word, bm->depth);
31 		if (ret < bm->depth)
32 			return true;
33 	}
34 
35 	return false;
36 }
37 
38 bool blk_mq_has_free_tags(struct blk_mq_tags *tags)
39 {
40 	if (!tags)
41 		return true;
42 
43 	return bt_has_free_tags(&tags->bitmap_tags);
44 }
45 
46 static inline void bt_index_inc(unsigned int *index)
47 {
48 	*index = (*index + 1) & (BT_WAIT_QUEUES - 1);
49 }
50 
51 /*
52  * If a previously inactive queue goes active, bump the active user count.
53  */
54 bool __blk_mq_tag_busy(struct blk_mq_hw_ctx *hctx)
55 {
56 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state) &&
57 	    !test_and_set_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
58 		atomic_inc(&hctx->tags->active_queues);
59 
60 	return true;
61 }
62 
63 /*
64  * Wakeup all potentially sleeping on normal (non-reserved) tags
65  */
66 static void blk_mq_tag_wakeup_all(struct blk_mq_tags *tags)
67 {
68 	struct blk_mq_bitmap_tags *bt;
69 	int i, wake_index;
70 
71 	bt = &tags->bitmap_tags;
72 	wake_index = bt->wake_index;
73 	for (i = 0; i < BT_WAIT_QUEUES; i++) {
74 		struct bt_wait_state *bs = &bt->bs[wake_index];
75 
76 		if (waitqueue_active(&bs->wait))
77 			wake_up(&bs->wait);
78 
79 		bt_index_inc(&wake_index);
80 	}
81 }
82 
83 /*
84  * If a previously busy queue goes inactive, potential waiters could now
85  * be allowed to queue. Wake them up and check.
86  */
87 void __blk_mq_tag_idle(struct blk_mq_hw_ctx *hctx)
88 {
89 	struct blk_mq_tags *tags = hctx->tags;
90 
91 	if (!test_and_clear_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
92 		return;
93 
94 	atomic_dec(&tags->active_queues);
95 
96 	blk_mq_tag_wakeup_all(tags);
97 }
98 
99 /*
100  * For shared tag users, we track the number of currently active users
101  * and attempt to provide a fair share of the tag depth for each of them.
102  */
103 static inline bool hctx_may_queue(struct blk_mq_hw_ctx *hctx,
104 				  struct blk_mq_bitmap_tags *bt)
105 {
106 	unsigned int depth, users;
107 
108 	if (!hctx || !(hctx->flags & BLK_MQ_F_TAG_SHARED))
109 		return true;
110 	if (!test_bit(BLK_MQ_S_TAG_ACTIVE, &hctx->state))
111 		return true;
112 
113 	/*
114 	 * Don't try dividing an ant
115 	 */
116 	if (bt->depth == 1)
117 		return true;
118 
119 	users = atomic_read(&hctx->tags->active_queues);
120 	if (!users)
121 		return true;
122 
123 	/*
124 	 * Allow at least some tags
125 	 */
126 	depth = max((bt->depth + users - 1) / users, 4U);
127 	return atomic_read(&hctx->nr_active) < depth;
128 }
129 
130 static int __bt_get_word(struct blk_align_bitmap *bm, unsigned int last_tag)
131 {
132 	int tag, org_last_tag, end;
133 
134 	org_last_tag = last_tag;
135 	end = bm->depth;
136 	do {
137 restart:
138 		tag = find_next_zero_bit(&bm->word, end, last_tag);
139 		if (unlikely(tag >= end)) {
140 			/*
141 			 * We started with an offset, start from 0 to
142 			 * exhaust the map.
143 			 */
144 			if (org_last_tag && last_tag) {
145 				end = last_tag;
146 				last_tag = 0;
147 				goto restart;
148 			}
149 			return -1;
150 		}
151 		last_tag = tag + 1;
152 	} while (test_and_set_bit_lock(tag, &bm->word));
153 
154 	return tag;
155 }
156 
157 /*
158  * Straight forward bitmap tag implementation, where each bit is a tag
159  * (cleared == free, and set == busy). The small twist is using per-cpu
160  * last_tag caches, which blk-mq stores in the blk_mq_ctx software queue
161  * contexts. This enables us to drastically limit the space searched,
162  * without dirtying an extra shared cacheline like we would if we stored
163  * the cache value inside the shared blk_mq_bitmap_tags structure. On top
164  * of that, each word of tags is in a separate cacheline. This means that
165  * multiple users will tend to stick to different cachelines, at least
166  * until the map is exhausted.
167  */
168 static int __bt_get(struct blk_mq_hw_ctx *hctx, struct blk_mq_bitmap_tags *bt,
169 		    unsigned int *tag_cache)
170 {
171 	unsigned int last_tag, org_last_tag;
172 	int index, i, tag;
173 
174 	if (!hctx_may_queue(hctx, bt))
175 		return -1;
176 
177 	last_tag = org_last_tag = *tag_cache;
178 	index = TAG_TO_INDEX(bt, last_tag);
179 
180 	for (i = 0; i < bt->map_nr; i++) {
181 		tag = __bt_get_word(&bt->map[index], TAG_TO_BIT(bt, last_tag));
182 		if (tag != -1) {
183 			tag += (index << bt->bits_per_word);
184 			goto done;
185 		}
186 
187 		last_tag = 0;
188 		if (++index >= bt->map_nr)
189 			index = 0;
190 	}
191 
192 	*tag_cache = 0;
193 	return -1;
194 
195 	/*
196 	 * Only update the cache from the allocation path, if we ended
197 	 * up using the specific cached tag.
198 	 */
199 done:
200 	if (tag == org_last_tag) {
201 		last_tag = tag + 1;
202 		if (last_tag >= bt->depth - 1)
203 			last_tag = 0;
204 
205 		*tag_cache = last_tag;
206 	}
207 
208 	return tag;
209 }
210 
211 static struct bt_wait_state *bt_wait_ptr(struct blk_mq_bitmap_tags *bt,
212 					 struct blk_mq_hw_ctx *hctx)
213 {
214 	struct bt_wait_state *bs;
215 
216 	if (!hctx)
217 		return &bt->bs[0];
218 
219 	bs = &bt->bs[hctx->wait_index];
220 	bt_index_inc(&hctx->wait_index);
221 	return bs;
222 }
223 
224 static int bt_get(struct blk_mq_alloc_data *data,
225 		struct blk_mq_bitmap_tags *bt,
226 		struct blk_mq_hw_ctx *hctx,
227 		unsigned int *last_tag)
228 {
229 	struct bt_wait_state *bs;
230 	DEFINE_WAIT(wait);
231 	int tag;
232 
233 	tag = __bt_get(hctx, bt, last_tag);
234 	if (tag != -1)
235 		return tag;
236 
237 	if (!(data->gfp & __GFP_WAIT))
238 		return -1;
239 
240 	bs = bt_wait_ptr(bt, hctx);
241 	do {
242 		bool was_empty;
243 
244 		was_empty = list_empty(&wait.task_list);
245 		prepare_to_wait(&bs->wait, &wait, TASK_UNINTERRUPTIBLE);
246 
247 		tag = __bt_get(hctx, bt, last_tag);
248 		if (tag != -1)
249 			break;
250 
251 		if (was_empty)
252 			atomic_set(&bs->wait_cnt, bt->wake_cnt);
253 
254 		blk_mq_put_ctx(data->ctx);
255 
256 		io_schedule();
257 
258 		data->ctx = blk_mq_get_ctx(data->q);
259 		data->hctx = data->q->mq_ops->map_queue(data->q,
260 				data->ctx->cpu);
261 		if (data->reserved) {
262 			bt = &data->hctx->tags->breserved_tags;
263 		} else {
264 			last_tag = &data->ctx->last_tag;
265 			hctx = data->hctx;
266 			bt = &hctx->tags->bitmap_tags;
267 		}
268 		finish_wait(&bs->wait, &wait);
269 		bs = bt_wait_ptr(bt, hctx);
270 	} while (1);
271 
272 	finish_wait(&bs->wait, &wait);
273 	return tag;
274 }
275 
276 static unsigned int __blk_mq_get_tag(struct blk_mq_alloc_data *data)
277 {
278 	int tag;
279 
280 	tag = bt_get(data, &data->hctx->tags->bitmap_tags, data->hctx,
281 			&data->ctx->last_tag);
282 	if (tag >= 0)
283 		return tag + data->hctx->tags->nr_reserved_tags;
284 
285 	return BLK_MQ_TAG_FAIL;
286 }
287 
288 static unsigned int __blk_mq_get_reserved_tag(struct blk_mq_alloc_data *data)
289 {
290 	int tag, zero = 0;
291 
292 	if (unlikely(!data->hctx->tags->nr_reserved_tags)) {
293 		WARN_ON_ONCE(1);
294 		return BLK_MQ_TAG_FAIL;
295 	}
296 
297 	tag = bt_get(data, &data->hctx->tags->breserved_tags, NULL, &zero);
298 	if (tag < 0)
299 		return BLK_MQ_TAG_FAIL;
300 
301 	return tag;
302 }
303 
304 unsigned int blk_mq_get_tag(struct blk_mq_alloc_data *data)
305 {
306 	if (!data->reserved)
307 		return __blk_mq_get_tag(data);
308 
309 	return __blk_mq_get_reserved_tag(data);
310 }
311 
312 static struct bt_wait_state *bt_wake_ptr(struct blk_mq_bitmap_tags *bt)
313 {
314 	int i, wake_index;
315 
316 	wake_index = bt->wake_index;
317 	for (i = 0; i < BT_WAIT_QUEUES; i++) {
318 		struct bt_wait_state *bs = &bt->bs[wake_index];
319 
320 		if (waitqueue_active(&bs->wait)) {
321 			if (wake_index != bt->wake_index)
322 				bt->wake_index = wake_index;
323 
324 			return bs;
325 		}
326 
327 		bt_index_inc(&wake_index);
328 	}
329 
330 	return NULL;
331 }
332 
333 static void bt_clear_tag(struct blk_mq_bitmap_tags *bt, unsigned int tag)
334 {
335 	const int index = TAG_TO_INDEX(bt, tag);
336 	struct bt_wait_state *bs;
337 
338 	/*
339 	 * The unlock memory barrier need to order access to req in free
340 	 * path and clearing tag bit
341 	 */
342 	clear_bit_unlock(TAG_TO_BIT(bt, tag), &bt->map[index].word);
343 
344 	bs = bt_wake_ptr(bt);
345 	if (bs && atomic_dec_and_test(&bs->wait_cnt)) {
346 		atomic_set(&bs->wait_cnt, bt->wake_cnt);
347 		bt_index_inc(&bt->wake_index);
348 		wake_up(&bs->wait);
349 	}
350 }
351 
352 static void __blk_mq_put_tag(struct blk_mq_tags *tags, unsigned int tag)
353 {
354 	BUG_ON(tag >= tags->nr_tags);
355 
356 	bt_clear_tag(&tags->bitmap_tags, tag);
357 }
358 
359 static void __blk_mq_put_reserved_tag(struct blk_mq_tags *tags,
360 				      unsigned int tag)
361 {
362 	BUG_ON(tag >= tags->nr_reserved_tags);
363 
364 	bt_clear_tag(&tags->breserved_tags, tag);
365 }
366 
367 void blk_mq_put_tag(struct blk_mq_hw_ctx *hctx, unsigned int tag,
368 		    unsigned int *last_tag)
369 {
370 	struct blk_mq_tags *tags = hctx->tags;
371 
372 	if (tag >= tags->nr_reserved_tags) {
373 		const int real_tag = tag - tags->nr_reserved_tags;
374 
375 		__blk_mq_put_tag(tags, real_tag);
376 		*last_tag = real_tag;
377 	} else
378 		__blk_mq_put_reserved_tag(tags, tag);
379 }
380 
381 static void bt_for_each_free(struct blk_mq_bitmap_tags *bt,
382 			     unsigned long *free_map, unsigned int off)
383 {
384 	int i;
385 
386 	for (i = 0; i < bt->map_nr; i++) {
387 		struct blk_align_bitmap *bm = &bt->map[i];
388 		int bit = 0;
389 
390 		do {
391 			bit = find_next_zero_bit(&bm->word, bm->depth, bit);
392 			if (bit >= bm->depth)
393 				break;
394 
395 			__set_bit(bit + off, free_map);
396 			bit++;
397 		} while (1);
398 
399 		off += (1 << bt->bits_per_word);
400 	}
401 }
402 
403 void blk_mq_tag_busy_iter(struct blk_mq_tags *tags,
404 			  void (*fn)(void *, unsigned long *), void *data)
405 {
406 	unsigned long *tag_map;
407 	size_t map_size;
408 
409 	map_size = ALIGN(tags->nr_tags, BITS_PER_LONG) / BITS_PER_LONG;
410 	tag_map = kzalloc(map_size * sizeof(unsigned long), GFP_ATOMIC);
411 	if (!tag_map)
412 		return;
413 
414 	bt_for_each_free(&tags->bitmap_tags, tag_map, tags->nr_reserved_tags);
415 	if (tags->nr_reserved_tags)
416 		bt_for_each_free(&tags->breserved_tags, tag_map, 0);
417 
418 	fn(data, tag_map);
419 	kfree(tag_map);
420 }
421 EXPORT_SYMBOL(blk_mq_tag_busy_iter);
422 
423 static unsigned int bt_unused_tags(struct blk_mq_bitmap_tags *bt)
424 {
425 	unsigned int i, used;
426 
427 	for (i = 0, used = 0; i < bt->map_nr; i++) {
428 		struct blk_align_bitmap *bm = &bt->map[i];
429 
430 		used += bitmap_weight(&bm->word, bm->depth);
431 	}
432 
433 	return bt->depth - used;
434 }
435 
436 static void bt_update_count(struct blk_mq_bitmap_tags *bt,
437 			    unsigned int depth)
438 {
439 	unsigned int tags_per_word = 1U << bt->bits_per_word;
440 	unsigned int map_depth = depth;
441 
442 	if (depth) {
443 		int i;
444 
445 		for (i = 0; i < bt->map_nr; i++) {
446 			bt->map[i].depth = min(map_depth, tags_per_word);
447 			map_depth -= bt->map[i].depth;
448 		}
449 	}
450 
451 	bt->wake_cnt = BT_WAIT_BATCH;
452 	if (bt->wake_cnt > depth / 4)
453 		bt->wake_cnt = max(1U, depth / 4);
454 
455 	bt->depth = depth;
456 }
457 
458 static int bt_alloc(struct blk_mq_bitmap_tags *bt, unsigned int depth,
459 			int node, bool reserved)
460 {
461 	int i;
462 
463 	bt->bits_per_word = ilog2(BITS_PER_LONG);
464 
465 	/*
466 	 * Depth can be zero for reserved tags, that's not a failure
467 	 * condition.
468 	 */
469 	if (depth) {
470 		unsigned int nr, tags_per_word;
471 
472 		tags_per_word = (1 << bt->bits_per_word);
473 
474 		/*
475 		 * If the tag space is small, shrink the number of tags
476 		 * per word so we spread over a few cachelines, at least.
477 		 * If less than 4 tags, just forget about it, it's not
478 		 * going to work optimally anyway.
479 		 */
480 		if (depth >= 4) {
481 			while (tags_per_word * 4 > depth) {
482 				bt->bits_per_word--;
483 				tags_per_word = (1 << bt->bits_per_word);
484 			}
485 		}
486 
487 		nr = ALIGN(depth, tags_per_word) / tags_per_word;
488 		bt->map = kzalloc_node(nr * sizeof(struct blk_align_bitmap),
489 						GFP_KERNEL, node);
490 		if (!bt->map)
491 			return -ENOMEM;
492 
493 		bt->map_nr = nr;
494 	}
495 
496 	bt->bs = kzalloc(BT_WAIT_QUEUES * sizeof(*bt->bs), GFP_KERNEL);
497 	if (!bt->bs) {
498 		kfree(bt->map);
499 		return -ENOMEM;
500 	}
501 
502 	for (i = 0; i < BT_WAIT_QUEUES; i++)
503 		init_waitqueue_head(&bt->bs[i].wait);
504 
505 	bt_update_count(bt, depth);
506 	return 0;
507 }
508 
509 static void bt_free(struct blk_mq_bitmap_tags *bt)
510 {
511 	kfree(bt->map);
512 	kfree(bt->bs);
513 }
514 
515 static struct blk_mq_tags *blk_mq_init_bitmap_tags(struct blk_mq_tags *tags,
516 						   int node)
517 {
518 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
519 
520 	if (bt_alloc(&tags->bitmap_tags, depth, node, false))
521 		goto enomem;
522 	if (bt_alloc(&tags->breserved_tags, tags->nr_reserved_tags, node, true))
523 		goto enomem;
524 
525 	return tags;
526 enomem:
527 	bt_free(&tags->bitmap_tags);
528 	kfree(tags);
529 	return NULL;
530 }
531 
532 struct blk_mq_tags *blk_mq_init_tags(unsigned int total_tags,
533 				     unsigned int reserved_tags, int node)
534 {
535 	struct blk_mq_tags *tags;
536 
537 	if (total_tags > BLK_MQ_TAG_MAX) {
538 		pr_err("blk-mq: tag depth too large\n");
539 		return NULL;
540 	}
541 
542 	tags = kzalloc_node(sizeof(*tags), GFP_KERNEL, node);
543 	if (!tags)
544 		return NULL;
545 
546 	tags->nr_tags = total_tags;
547 	tags->nr_reserved_tags = reserved_tags;
548 
549 	return blk_mq_init_bitmap_tags(tags, node);
550 }
551 
552 void blk_mq_free_tags(struct blk_mq_tags *tags)
553 {
554 	bt_free(&tags->bitmap_tags);
555 	bt_free(&tags->breserved_tags);
556 	kfree(tags);
557 }
558 
559 void blk_mq_tag_init_last_tag(struct blk_mq_tags *tags, unsigned int *tag)
560 {
561 	unsigned int depth = tags->nr_tags - tags->nr_reserved_tags;
562 
563 	*tag = prandom_u32() % depth;
564 }
565 
566 int blk_mq_tag_update_depth(struct blk_mq_tags *tags, unsigned int tdepth)
567 {
568 	tdepth -= tags->nr_reserved_tags;
569 	if (tdepth > tags->nr_tags)
570 		return -EINVAL;
571 
572 	/*
573 	 * Don't need (or can't) update reserved tags here, they remain
574 	 * static and should never need resizing.
575 	 */
576 	bt_update_count(&tags->bitmap_tags, tdepth);
577 	blk_mq_tag_wakeup_all(tags);
578 	return 0;
579 }
580 
581 ssize_t blk_mq_tag_sysfs_show(struct blk_mq_tags *tags, char *page)
582 {
583 	char *orig_page = page;
584 	unsigned int free, res;
585 
586 	if (!tags)
587 		return 0;
588 
589 	page += sprintf(page, "nr_tags=%u, reserved_tags=%u, "
590 			"bits_per_word=%u\n",
591 			tags->nr_tags, tags->nr_reserved_tags,
592 			tags->bitmap_tags.bits_per_word);
593 
594 	free = bt_unused_tags(&tags->bitmap_tags);
595 	res = bt_unused_tags(&tags->breserved_tags);
596 
597 	page += sprintf(page, "nr_free=%u, nr_reserved=%u\n", free, res);
598 	page += sprintf(page, "active_queues=%u\n", atomic_read(&tags->active_queues));
599 
600 	return page - orig_page;
601 }
602